Multi-scale curvature product for robust image corner detection in curvature scale space

The paper proposes a new corner detector based on scale multiplication. The algorithm starts with extracting the contour of the object of interest, and then computes the curvature of this contour with Gaussian derivative filters at various scales. Local extremes of the product of the curvatures at different scales are reported as corners when the value of the product exceeds a threshold. The proposed detector is based on the well-known curvature scale-space, but improves on it in two ways: first, since the finest scale is part of the scale product, there is no need for coarse-to-fine corner tracking. Second, since many scales are involved, false positive/negative detections are unlikely even with a single threshold. Finally, the comparison between the proposed approach and other corner detectors shows that our approach is more competitive with respect to CCN and ACU criteria under similarity and affine transforms. Moreover, a number of experiments also illustrate that the scale product corner detection has more robustness for noise.